IJCSNS International Journal of Computer Science and Network Security, VOL.11 No.2, Feb. 2011
143
Number Plate Recognition for Indian Vehicles
M. M. Shidore†, S. P. Narote††
†
Vishwakarma Institute of Technology, University of Pune, Pune, India
††
Sinhgad College of Engineering, University of Pune, Pune, India
Abstract: This paper presents Automatic Number Plate
extraction, character segmentation and recognition for
Indian vehicles. In India, number plate models are not
followed strictly. Characters on plate are in different
Indian languages, as well as in English. Due to variations
in the representation of number plates, vehicle number
plate extraction, character segmentation and recognition
are crucial. We present the number plate extraction,
character segmentation and recognition work, with english
characters. Number plate extraction is done using Sobel
filter, morphological operations and connected component
analysis. Character segmentation is done by using
connected component and vertical projection analysis.
Character recognition is carried out using Support Vector
machine (SVM). The segmentation accuracy is 80% and
recognition rate is 79.84 %.
Key words:
Vehicle number plate, Number plate extraction, character
segmentation, character recognition.
1. Introduction
Automatic Number Plate Recognition (ANPR) system is
an important technique, used in Intelligent Transportation
System. ANPR is an advanced machine vision technology
used to identify vehicles by their number plates without
direct human intervention. It is an important area of
research due to its many applications. The development of
Intelligent Transportation System provides the data of
vehicle numbers which can be used in follow up, analyses
and monitoring. ANPR is important in the area of traffic
problems, highway toll collection, borders and custom
security, premises where high security is needed, like
Parliament, Legislative Assembly, and so on. The
complexity of automatic number plate recognition work
varies throughout the world. For the standard number plate,
ANPR system is easier to read and recognize. In India this
task becomes much difficult due to variation in plate
model.
The ANPR work is generally framed into the steps:
Number plate extraction, character segmentation and
character recognition. From the entire input image, only
the number plate is detected and processed further in the
next step of character segmentation. In character
segmentation phase each and every character is isolated
and segmented. Based on the selection of prominent
Manuscript received February 5, 2011
Manuscript revised February 20, 2011
features of characters, each character is recognized, in the
character recognition phase. Extraction of number plate is
difficult task, essentially due to: Number plates generally
occupy a small portion of whole image; difference in
number plate formats, and influence of environmental
factors. This step affects the accuracy of character
segmentation and recognition work. Different techniques
are developed for number plate extraction.
Hao Chen et al [1] proposed the method in two steps.
First, several candidates based on texture information
similar to number plate are extracted. In the second step,
auto-correlation based binary image and projection
algorithm are used to verify the true candidate plate. Gisu
Heo [2] developed number plate detection algorithm using
group of lines forming rectangle at the plate boundary.
Followed by this step is the vertical edge density algorithm
to find out the plate area. Ozbay et al [3] devised smearing
algorithm to locate the number plate. In Korean license
plate extraction, Mei Yu et al [4] proposed vertical edge
detection followed by size, shape filter for edge area and
edge matching technique based on plate model. Farhad
Faradji et al [5] first used Sobel vertical edge detection on
the image. Next, vertical projection analysis was used to
locate plate area. Fake candidates were deleted using
compact factor, which evaluated the densest vertical edge
area declaring true number plate.
Every character on detected number plates is
segmented in character segmentation step. Segmentation
techniques based on projection analysis, Hough transform,
region growing are proposed in the literature. Xinagjian
He et al [6] used horizontal and vertical projection analysis
for character segmentation.
Yuangang Zhang et al [7] implemented character
segmentation using Hough Transform. In this, horizontal
edges of the plate area were decided initially, using Hough
Transform, which helped to segment the characters with
the large rotation. Characters were segmented using
vertical projection analysis based on the prior knowledge
of the plate model. Feng Yang et al [8] developed region
growing algorithm for character segmentation. Shen Zheng
Wang et al [9] used connected component analysis for
character segmentation.
Fundamental issue in character recognition is shape
analysis. For the recognition of segmented characters,
algorithms use Pattern/ Template matching, Computational
Intelligence, statistical classifiers.
144
IJCSNS International Journal of Computer Science and Network Security, VOL.11 No.2, February 2011
ANPR system consists of three modules; 1) Extraction
of number plate, 2) Character segmentation, and 3)
Character recognition. In this paper ANPR wok for Indian
car is presented. Images are taken out with different
illumination conditions, different background and
orientation. Histogram equalization, median filter are used
which take care of lighting and contrast problem. Sobel
vertical edge detection and morphology is employed to
locate the number plate. Projection analysis is used to
segment the characters present on the plate. For
recognition work SVM is used.
The rest of the paper is organized as follows: Section 2
explains the proposed algorithm for number plate
extraction; Section 3 explains the algorithm used for
character segmentation; Section 4 describes the number
plate recognition algorithm using SVM. Experimental
results are demonstrated in Section 5. Section 6 concludes
the paper.
2.3 Candidate Plate Area Detection
2. Number Plate Extraction
2.4 True Number Plate Extraction
Number plate extraction is the key step in ANPR system,
which influences the accuracy of the system significantly.
The goal of this phase, given an input image, is to produce
a number of candidate regions, with high probability of
containing number plate and validate for true number plate.
After the detection of candidate number plate area,
Bounding Box analysis is used to extract plate area from
the original image. From the Bounding Box analysis,
respective row and column indices of plate area are found
out. Once the indices of number plate are known, the
number plate is extracted from original gray scale image.
The result is as shown in Fig. 2.
2.1 Image Acquisition and Preprocessing
In this system a high resolution digital camera is used to
acquire an image. Images are taken in different
background, illumination conditions, and at various
distances from the camera to vehicle. Images are resized to
(1024 X 768). All the processing steps are executed on
gray scale image. Preprocessing is mainly used to enhance
the processing speed, improve the contrast of the image,
and to reduce the noise in the image. In order to reduce the
problem of low quality and low contrast in car images,
images are enhanced by using histogram equalization on
gray scale image.
2.2 Vertical Edge detection
The characters on number plate region contain abundant
edges as compared to background area. This feature is
employed for locating the candidate plate area from the
input image. Sobel vertical edge detection is used to find
out the regions which have high pixel variance value [10].
To extract candidate number plate area from the entire
image, threshold is used to select rows which are having
particular white pixel density. Fig. 1 shows the result of
effect of using, Sobel vertical edge detection and
threshold.
(a)
(b)
(c)
Fig. 1: (a) Gray scale image (b) Sobel vertical edge detection (c) Effect of
threshold
Morphological operations aim to remove unrelated objects
in the image. Dilation and erosion are used to extract
candidate plate areas from the entire image. Sometimes
background areas may also get declared as candidate plate.
Hence to remove the fake candidates, plate validation is
done using the aspect ratio of the plate and horizontal cuts
[11] in the number plate.
Fig. 2: Extracted number plate
3. Character Segmentation
Character isolation from the number plate region is the
important step in ANPR system, which influences the
accuracy of character recognition significantly. The goal of
this phase, given the number plate image, is to segment all
the characters, without loosing features of the characters.
This phase consists of the sequence of operations as,
character region enhancement, connected component
analysis and projection analysis.
3.1 Character Region Enhancement
To enhance the character region, if histogram equalization
or gray level scaling is used, noise may also be enhanced.
For the correct character segmentation, only the character
pixels need to be enhanced and the background pixels need
to be weakened. For this ‘gray threshold’ [10] is
implemented. Fig. 3 shows thresholding effect.
Fig. 3: Binarized number plate
3.2 Connected Component Analysis
IJCSNS International Journal of Computer Science and Network Security, VOL.11 No.2, February 2011
To remove the noise, other than characters on the plate,
connected component analysis is used. Each labeled
matrix of 8-connectivity pixels is evaluated based on the
area threshold. The result of connected component
analysis and noise removal is as shown in Fig. 4.
(a)
(b)
Fig. 4: (a) Connected component analysis (b) Noise removed
145
utilization time, particularly when the features are
redundant.
Initially, the centroid of the character image is
determined.
With respect to centroid, number of
transitions along the axes, 0 to 1 and 1 to 0, up to the
boundary of character are counted. Transitions are
specified for axes with predetermined angles. Fig. 6 shows
the nature of axes towards boundary, counting number of
transitions. By observing strokes of each character, 13
different angles are decided to count the transitions.
3.3 Vertical Projection Analysis
Character segmentation is implemented to isolate each
character on the number plate. Vertical projection analysis
is used to find the gaps between the characters. The
characters are segmented based on the number of valleys
in projection. Row and column indices of each character
are recorded, and it is extracted from the original gray
scale number plate. The vertical projection and segmented
characters are shown in Fig. 5. This proposed work
isolates every letter and digit on the number plate oriented
horizontally in one row, along the width of number plate.
Fig. 5: Vertical projection analysis and respective segmented characters
4. Character Recognition
The character recognition phase consists of two steps: 1)
Character normalization and feature extraction, 2)
Character classification using Support Vector Machine
[12].
4.1 Character Normalization
Segmented characters have very much variation in size. In
this phase, all the characters are normalized to predefined
height (Vertical Length) in pixel. As the characters always
have variable width (Horizontal Length), each character
image is normalized to a size of 32 X 32, by image
mapping technique.
4.2 Feature Extraction
The goal of feature vector is to define distinguishing
features of the characters. Selecting the most relevant
feature of each character can not only facilitate data
visualization and data understanding, but also reduce the
measurement, storage requirements, training and
Fig. 6: Input character image overlaid with nature of axes
4.3 Number Recognition
The character recognition work is carried out using SVM
as classifier. The car number plate consists of combination
of 0 to 9 digits and alphabets of A to Z. All the digits as
well as alphabets are assigned a class label. For multi-class
classification, either one-against-one approach is used or
one-against-all is preferred. In the proposed work
recognition is achieved using one-against-all approach
[13]. The polynomial kernel is used for classification. The
database consists of 250 images. 60 % of the data is used
for training and remaining 40% is used for testing.
5. Experimental Results and Performance
Analysis
A database consists of different sized JPEG colored
images. Images are resized to 1024 X 768. Total 250
images are used to test the algorithm. The images are
taken with different background as well as illumination
conditions. Experiments show that the algorithm has good
performance on number plate extraction, and character
segmentation work. It can deal the images correctly, with
noise, illumination variance, and rotation to ±50. This work
is implemented using MATLAB.
Table 1 illustrates number plate extraction and
character segmentation success rate.
Table1 Performance of number plate extraction and character
segmentation.
Total No.
Success
Algorithm
of Images
Rate (%)
Number
Plate
250
85
Extraction
Character
214
80
Segmentation
Deep shadows and reflections have an impact on number
plate extraction work. Because of uneven illumination,
stained number plates, true number plates could not get
IJCSNS International Journal of Computer Science and Network Security, VOL.11 No.2, February 2011
146
correctly extracted. Failure in character segmentation was
mainly because of merging of characters on number plate,
stained number plates, orientation of the image and poor
illumination.
Character recognition work is done on 10 digits (0 to
9) and 26 alphabets (A to Z). The recognition rate achieved
using ‘polynomial’ kernel is 79.84 %. The false
recognition is due to similarity in the character shape, e.g.
6 and B, 5 and S etc. Recognition can be increased if two
groups of SVM’s are designed separately, one for digits
and other for alphabets.
[10]
[11]
[12]
[13]
Rafel C. Gonzalez, Richard E. Woods, “Digital Image
Processing,” 2nd edition, Prentice Hall, Inc., 2002.
Tran Dur Duan, Duong Anh Duc, Tran Le, Hong Du,
“Combining Hough transform and contour algorithm
for detecting vehicle license plate”, Proc. of
International Symposium Intelligent Multimedia, Video
and Speech Processing , 2004, pp. 747-750.
Christopher J. C. Burges, “A tutorial on Support Vector
Machine for pattern recognition,” Bell Laboratories,
Lucent Technologies, 1998.
S. R. Gunn, “Support Vector Machines for
classification and regression, technical report,”
University of Southampton, 1998.
6. Conclusions
An algorithm for vehicle number plate extraction,
character segmentation and recognition is presented.
Database of the image consists of images with different
size, background, illumination, camera angle, distance etc.
The experimental results show that, number plates
are extracted faithfully based on vertical edge detection
and connected component algorithm, with the success rate
of 85%. Character segmentation phase using connected
component analysis and vertical projection analysis works
well with the success rate of 80%. The success rate
achieved for character recognition is 79.84%.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
Hao Chen, Jisheng Ren, Huachun Tan, Jianqun Wang,
“ A novel method for license plate localization”, 4th
Proc. of ICIG 2007, pp. 604-609.
Gisu Heo, Minwoo Kim, Insook Jung, Duk Ryong Lee,
Il Seok Oh, “Extraction of car license plate regions
using line grouping and edge density methods”,
International Symposium on Information Technology
Convergence, 2007, pp. 37-42.
Serkan Ozbay, Ergun Ercelebi, “Automatic vehicle
identification by plate recognition”, Proc. of PWASET,
vol. 9, no. 4, 2005, pp. 222-225.
Mei Yu and Yong Deak Kim,
“An approach to
Korean license plate recognition based on vertical edge
matching”, IEEE International Conference on System,
Man and Cybernetics, 2000, vol.4, pp. 2975-2980.
Farhad Faradji, Amir Hossein Rezaie, Majid Ziaratban,
“A Morphological based License Plate Location”, ICIP,
2007, pp. I 57-I 60.
Xiangjian He et al, “Segmentation of characters on car
license plates”, 10th Workshop on Multimedia Signal
Processing, Oct. 2008, pp. 399-402.
Yungang Zhang, Changshui Zhang, “A New algorithm
for character segmentation of license plate”, Proc. of
IEEE Intelligent Vehicles Symposium, 2003, pp.
106-109.
Feng Yang, Zheng Ma, Mei Xie, “A Novel approach
for license plate character segmentation”, ICIEA, 2006,
pp.1-6.
Shen-Zheng Wang, His-Jian Lee, “Detection and
recognition of license plate characters with different
appearances,” Proc. of 16th International Conference on
Pattern Recognition, vol.3, 2003, pp. 979-983.
Mrunal M. Shidore was born in
India in 1964. She received the
B.E. (Electronics) degree from
University of Pune, Pune in 2003.
She is currently working as a
Lecturer
in
Electronics
Engineering,
with
the
Vishwakarma
Institute
of
Technology, Pune. Her research and teaching interests
include image processing.
Sandipan P Narote was born in
India in 1975. He received the
B.E.(E&TC), M.E.(Elect.) from
Dr.
Babasaheb
Ambedkar
University
Aurangabad
and
Ph.D.(E & TC) from Swami
Ramanand Tirth Marathwada University Nanded in 2000,
2002 and 2010, respectively. He is currently a Professor of
Electronics and Telecommunication Engineering with the
Sinhgad College of Engineering, Pune, India.
His
research and teaching interests include signal processing,
image processing and biometrics.